anyangsong/MAGA-ROLE-80

This is the repository of system role prompts used by our MAGA, inspired by ROLE-88.

Current research in the field of roleplaying mainly focuses on providing fine-grained information such as the attributes, relations, and biographical history of a specific role, with the expectation that large language models (LLMs) can make decisions similar to that specific role on certain tasks (e.g., multiple-choice questions in psychology). However, when a large amount of background information is introduced, the model cannot process it effectively, and the degree of anthropomorphism deteriorates significantly. On this basis, attempts are made to improve the effect of fitting a specific role, but the fitting is not targeted at the general set of humans. This is a characteristic and challenge of this research field.

Coarse-grained anthropomorphism, especially some simple and clear instructions, can achieve good role-playing results as long as they are specified in the system role. This has become a consensus in practical applications, including but not limited to some entertainment-oriented communities for LLM writing creation. There are few studies on coarse-grained anthropomorphism, and in fact, most of them are still in the position paper stage rather than more specific research tasks.

Inspired by ROLE-88, we manually compiled MAGA-ROLE-80, which includes 80 prompts each in Chinese and English.

A single role is obviously biased. If it is used in machine-generated texts for training AI detectors, it will obviously degenerate into style detection, such as detecting whether a text is written in an American style.

MAGA-ROLE-80 covers many major, easily recognizable coarse-grained abstract social occupations and identities of humans. It uses dozens of enhanced languages with different styles, starting from local details to cover the global, making the dataset more anthropomorphic overall and better aligned with human characteristics.

To facilitate better role-playing of these characters, nearly every prompt in MAGA-ROLE-80 further supplements explanations of tone, intonation, and lexical characteristics. After all, what we really care about is the characteristics and style of the text itself.

It should be noted, however, that approximately 20% of the prompts are quite special, depicting well-known social celebrities and famous characters from novels, film and television dramas, and games. These roles have extremely distinct style characteristics, covering more like individual points rather than local areas, which clearly deviates from the overall anthropomorphism objective. Nevertheless, as a minor part, these roles with distinct characteristics are very helpful for expanding the diversity of the dataset and will not have a fundamental impact on the main components. After all, mere diversity is also of great value.

Load the dataset

To load the dataset, install the library datasets with pip install datasets. Then,

from datasets import load_dataset
maga-role-80 = load_dataset("anyangsong/MAGA-ROLE-80")

Citation

If you find MAGA useful for your research and applications, please cite using the Bibtex:

@misc{song2026maga,
      title={MAGA-Bench: Machine-Augment-Generated Text via Alignment Detection Benchmark}, 
      author={Anyang Song and Ying Cheng and Yiqian Xu and Rui Feng},
      year={2026},
      eprint={2601.04633},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.04633}, 
}